KR102320755B1 - On-device machine learning platform - Google Patents

Abstract

The present disclosure provides systems and methods for on-device machine learning. In particular, this disclosure relates to on-device machine learning platforms and related technologies that enable on-device prediction, training, example collection, and/or other machine learning tasks or functions. The on-device machine learning platform may include a context provider that securely inserts context features into the collected training examples and/or client-supplied input data used to generate predictions/inferences. Thus, an on-device machine learning platform can use centralized training example collection, model training, and use of machine learning models as a service to applications or other clients.

Description

On-device machine learning platform

The present invention relates generally to machine learning. More specifically, the present invention relates to on-device machine learning platforms and related technologies that enable on-device prediction, training, example collection, and/or other machine learning tasks or functions.

In recent years, machine learning is increasingly used to provide improved services to users of computing devices. In particular, many applications or other computing programs or systems rely on one or more machine learning models to generate inferences based on input data related to programs, devices, and/or users. The application(s) may use inference to perform or affect any type of task or service.

One conventional training approach for solving machine learning problems involves collecting a plurality of training examples from a plurality of computing devices (eg, user devices such as smart phones) in a centralized location (eg, a server device). include Thereafter, a machine learning model may be trained at a centralized location based on the collected training examples.

Also, in some cases, the trained model may be stored in a centralized location. To receive inference from the model, the user computing device transmits input data to the server computing device via the network, until the server device implements the machine learning model to generate inference(s) based on the transmitted data. It waits and then receives the inference(s) from the server computing device back over the network.

In such a scenario, training example(s) and/or inference(s) are required to be transmitted between the user computing device and the server computing device over a network. Such network transmissions represent a data security risk as data transmitted over the network may be susceptible to interception. In addition, these network transmissions increase network traffic, which can slow down communication speeds. In addition, latency related to data transmission/reception through the network may cause a delay in providing an application service.

More recently, specific applications include machine learning models that are stored within applications and implemented by applications on user devices. However, this architecture suffers from implementation and resource intensive issues. For example, in such a scenario, the application requires storing, managing, training, and/or implementing one or more machine learning models. Including the model and its supporting services within the application itself increases the data size of the application and takes up more memory space.

Machine learning within an application may also require more frequent application updates. For example, an application may be required to be updated as the underlying machine learning engine is updated or evolved. An application update may undesirably require the user to use and stop the network when the update is downloaded and installed.

Additionally, application development can be complicated by machine learning within the application as additional services must be built into the application itself. As such, developers may have to learn and maintain the complexities of other machine learning engines.

Aspects and advantages of embodiments of the present disclosure will be set forth in part in the description that follows, may be learned from the description, or may be learned through practice of the embodiments.

One illustrative aspect of the present disclosure relates to a computing device. A computing device includes one or more processors and one or more non-transitory computer-readable media. The one or more non-transitory computer-readable media may include one or more applications implemented by one or more processors; one or more machine learning models; store instructions that, when executed by one or more processors, cause the computing device to implement an on-device machine learning platform that performs operations. The operations include receiving input data from a first one of the one or more applications via the predictive application programming interface. The operations include using at least a first machine learning model of the one or more machine learning models to generate at least one inference based at least in part on the input data. The operations include providing at least one inference generated by the first machine learning model to the first application via the predictive application programming interface.

Another illustrative aspect of the present disclosure is in one or more non-transitory computer-readable media that collectively store instructions that, when executed by one or more processors, cause a computing device to implement an on-device machine learning platform that performs operations. it's about The operations include receiving input data from a first one of the one or more applications stored on the computing device via the predictive application programming interface. The operations include using at least a first one of the one or more machine learning models stored on the computing device to generate at least one inference based at least in part on the input data. The operations include providing at least one inference generated by the first machine learning model to the first application via the predictive application programming interface.

Another exemplary aspect of the present disclosure relates to a computer-implemented method. The method includes receiving, by the computing device via an aggregation application programming interface, a new training example from a first one of a plurality of applications stored on the computing device. The method includes storing, by the computing device, new training examples in a centralized example database of the computing device. The method comprises from a first application via a training application programming interface to retrain, by the computing device, a first machine learning model stored by the computing device based at least in part on one or more training examples stored by the centralized example database. receiving an instruction. The method includes, in response to the instruction, causing, by the computing device, the first machine learning model to be retrained based at least in part on one or more training examples stored by the centralized example database. The method includes receiving, by a computing device, input data from a first application via a predictive application programming interface. The method includes using the first machine learning model to generate at least one inference based at least in part on input data. The method includes providing at least one inference generated by the first machine learning model to a first application via a predictive application programming interface.

Other aspects of the disclosure relate to various systems, apparatus, non-transitory computer-readable media, user interfaces, and electronic devices.

These and other features, aspects and advantages of various embodiments of the present disclosure will be better understood with reference to the following description and appended claims. BRIEF DESCRIPTION OF THE DRAWINGS The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate exemplary embodiments of the present disclosure, and together with the description serve to explain related principles.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A detailed description of embodiments, which is intended for those skilled in the art, is presented herein, and reference is made to the accompanying drawings.
1 depicts a block diagram of an exemplary computing system including an on-device machine learning platform in accordance with an exemplary embodiment of the present disclosure.
2 depicts a graphical diagram of an exemplary machine learning model deployment in accordance with an exemplary embodiment of the present disclosure.
3 shows a graphical diagram of an exemplary personalized and federated learning data flow in accordance with an exemplary embodiment of the present disclosure.
4 depicts a block diagram of an exemplary on-device machine learning platform in accordance with an exemplary embodiment of the present disclosure.
5A and 5B show block diagrams of an example machine learning platform for embedding context features in accordance with example embodiments of the present disclosure.
6A and 6B show block diagrams of an exemplary device for performing model training in accordance with an exemplary embodiment of the present disclosure.
7 shows a graphical diagram of an exemplary federated learning process in accordance with an exemplary embodiment of the present disclosure.
8 depicts a flow diagram of an exemplary method for generating inferences using a machine learning model in accordance with an exemplary embodiment of the present disclosure.
9 depicts a flow diagram of an exemplary method of collecting training examples for performing machine learning in accordance with an exemplary embodiment of the present disclosure.
10 depicts a flow diagram of an exemplary method of training a machine learning model in accordance with an exemplary embodiment of the present disclosure.

Generally, this disclosure relates to systems and methods for on-device machine learning. In particular, this disclosure relates to on-device machine learning platforms and related technologies that enable on-device prediction, training, example collection, and/or other machine learning tasks or functions, collectively referred to as “machine learning capabilities”. can be

The on-device machine learning platform may be in the form of one or more computer programs stored locally on a computing device or terminal (eg, a smart phone or tablet), which, when executed by a user device or terminal, may be in the form of one or more locally stored and perform machine learning management operations that enable execution of on-device machine learning functions on behalf of applications, routines, or other local clients. At least some of the on-device machine learning functions may be performed using one or more machine learning engines implemented locally on the computing device or terminal. The performance of on-device machine learning functions on behalf of one or more locally stored applications or routines (which may be referred to as “clients”) may interact with the on-device machine learning platform through one or more application programming interfaces (APIs). It can be provided to the corresponding client as a centralized service.

Additionally, in some implementations, the on-device machine learning platform may include a context provider that securely embeds contextual features into client-supplied input data and/or collected training examples used to generate predictions/inferences. Thus, an on-device machine learning platform can use centralized training example collection, model training, and use of machine learning models as a service to applications or other clients.

More specifically, a computing device, such as, for example, a mobile computing device (eg, a smart phone) may store or include one or more applications (eg, a mobile application). The computing device may also include and implement an on-device machine learning platform and one or more machine learning models. For example, the machine learning model may be stored by the device in a centralized model layer managed by the platform.

According to one aspect of the present disclosure, an application communicates with an on-device machine learning platform via an API (which may be referred to as a "prediction API") to provide input data and based on the input data from one or more machine learning models. predictions can be obtained. As an example, in some implementations, given a prediction plan (eg, instructions to run a model to obtain inference/prediction) and a Uniform Resource Identifier (URI) for model parameters, an on-device machine learning platform may download the URI content (eg, prediction plans and parameters) and obtain one or more inferences/predictions (eg, by interacting with a machine learning engine to implement the model by the engine). Additionally, the platform may cache content for use in subsequent prediction requests.

Thus, on-device machine learning models can be accessed by applications by communicating with the on-device machine learning platform through a client/service relationship. In particular, in some implementations, the machine learning platform may be a standalone multi-tenant service that may be referenced by an application. Thus, a given application does not need to store, manage, train and/or implement the machine learning model(s), but may instead communicate with the on-device machine learning platform to request and receive inferences from the model(s).

According to another aspect of the present disclosure, the computing device may further include a centralized example database that stores training examples received from one or more applications. In particular, the on-device machine learning platform may receive training examples from applications via an API (which may be referred to as an "ingestion API") and manage storage of examples in a centralized example database. For example, each application that is a client or tenant of the platform may have its own example collection(s) stored within a centralized example database, which collection(s) may be supplemented and/or managed in an online manner.

In some implementations, the on-device machine learning platform may cause storage of each training example received from the application (eg, within its corresponding collection) according to one or more optional parameters associated with the application providing the training example. have. As an example, the optional parameters may include a time-to-live (TTL) parameter that defines a period of time during which the training example is stored (eg, deleted thereafter). In some implementations, optional parameters may be predefined and/or adjusted via commands provided to the platform via the aggregation API.

According to another aspect of the present disclosure, in some implementations, the on-device machine learning platform can safely insert contextual features that describe the context associated with the computing device into the training example. For example, upon receiving training examples from an application, the context provider component of the on-device platform may determine one or more context characteristics and store these context characteristics along with the training examples in a centralized example database. For example, context features and data provided in a new training example may be stored as a single database item. Specific contextual characteristics determined and then inserted or otherwise associated and/or stored with a training example received from a particular application may be specified by optional parameters for that particular application. As described above, these optional features can be pre-defined or tuned via the ingest API. Thus, an application can control (eg, by defining optional parameters) which context characteristics or context types are inserted into its training examples.

In some implementations, context features are inserted on the service side so that the context features do not have to be available directly to the application. In particular, in some implementations, the centralized database of examples is not directly accessible by one or more applications, so that contextual information stored with a particular training example cannot be accessed even by the application that provided the training example.

In some implementations, context characteristics may be grouped or otherwise classified according to a number of different context types. In general, each context type may specify or contain a well-known name and a set of context characteristics of the well-known type. One example of a context type is device information including exemplary context characteristics such as audio status, network status, power connection, and the like.

In some implementations, the context provider requests the inserted value for a given context characteristic from the device (eg, the device's context manager) at insertion time/point. Alternatively or additionally, a context provider may register as a listener for one or more context updates, and may maintain a context feature cache of current values for context features based on one or more context updates. Then, when the context feature is inserted, the context provider accesses the context feature cache and inserts the current value held in the cache for that particular context feature.

In addition to or alternatively to inserting the context feature into the training example at storage time, the context provider may also perform the inserting of the context feature at inference time. Similar to the process described above for gathering training examples, in particular, when a particular application or other client requests that an inference be generated based on some client-supplied input data (e.g., via a prediction API), the context provider can You can insert or provide supplemental contextual features for input into the corresponding machine learning model along with the data. Accordingly, inferences may be made based at least in part on context information in addition to client-provided input data, which may help to improve the accuracy of the inferences.

In addition to the above description, the user can determine when and when a system, program, or feature described herein may enable the collection of user information (eg, training examples and contextual features), and when and when the user can obtain content from the server. Alternatively, a control may be provided that allows the user to make a choice as to when the communication is sent. In addition, certain data may be processed in one or more ways prior to storage or use so that personally identifiable information is removed. For example, the user's identity may be processed such that personally identifiable information about the user cannot be determined, or the user's geographic location may be determined from where the location information was obtained such that a specific location of the user cannot be determined (city, zip code or state level). ) can be generalized. Thus, the user can control what information is collected about the user, how that information is used, and what information is provided to the user.

According to another aspect, the context provider may have the right to access only certain context features or context types (eg, as defined or controlled by the device user), so that a particular application or other client may have the client privileges. control can be performed. In particular, in some implementations, an on-device machine learning platform or other device component may maintain a mapping of which clients have permission to access which context types. When a context feature is to be inserted (e.g., as a training example for storage or to supplement input data provided by the client at inference time), the context provider can provide a corresponding application or You can check the permission status of other clients. For example, the permission status for a particular application and context type may describe whether that application has permission to access that context type. The context provider will prevent (even indirectly) from accessing (even indirectly) context features/types that the application does not have access to, by inserting only the context functions contained in the context type that the application has permission to access.

Similar to the validity period option parameter described above, in some implementations, each context feature may have an expiration period associated with or assigned to it. This expiry period information may be associated with each training example with context feature(s) included. In some implementations, at the end of an expiration period for a particular context feature provided in a particular training example, the value for that context feature may be deleted or removed from that training example. Alternatively, the entire training example may be deleted or removed.

Further, in some implementations, in response to a change to a privilege state for a particular application or other client for a particular context characteristic or context type, the on-device platform retrieves these contexts associated with training examples related to the particular application from a centralized example database. Any value or item for a feature(s) or type(s) may be deleted. Also, in some implementations, the corresponding model(s) may be retrained on the remaining data after deletion of the context feature value(s).

According to another aspect of the present disclosure, an application communicates with an on-device machine learning platform via an API (also referred to as a "training API") to retrain or retrain a machine learning model based on training examples stored in a centralized database of examples. It may cause an update. As an example, in some implementations, given a URI to a training plan (eg, an instruction for training a model), the on-device machine learning platform is configured based on previously collected examples (eg, by the engine You can run training on a model (by interacting with a machine learning engine to trigger training of the model). For example, training may be performed in the background at scheduled times and/or when the device is idle.

After retraining the model, the retrained model can be used to provide inferences as described elsewhere in this disclosure. In general, such inferences will have higher accuracy because the model has been retrained on user-specific data. Thus, an on-device machine learning platform may enable centralized example data collection and corresponding personalization of machine learning models as a service to applications or other clients.

According to another aspect, in some implementations, the machine learning platform may upload logs or other updates regarding the machine learning model to the cloud for detailed analysis of machine learning metrics. As an example, in some implementations, the on-device platform may determine an update describing a parameter of a retrained machine learning model or changes to a parameter of the machine learning model that occurred during retraining of the model. The platform may send updates to a central server computing device (eg, the “cloud”) for integration with other updates provided by other computing devices. Thus, the platform may engage in a process known as “federated learning,” in which the device determines local updates to the model based on locally stored data and then aggregates those local updates for aggregation (e.g., privacy protection and in a communication-efficient way) to create a global update to that model.

According to another aspect, in some implementations, to protect applications from each other, each application has its own enclave for specific functions (eg, for all functions) provided by the on-device platform. can have For example, the platform may authenticate the application before the interface for accessing the platform is returned to the application through a factory. The returned interface can then represent the only view of the application enclave on the platform. In an example implementation of this process, when an application connects to the platform's API, the application may provide a signed package token that verifies the identity of the application. An application cannot obtain an API interface without passing this authentication.

According to another aspect of the present disclosure, in some implementations, the on-device machine learning platform may be completely abstracted from the underlying machine learning engine. For example, the machine learning engine may be a tensorflow engine, a neural network library, or other engine capable of implementing machine learning models for inference and/or training. This abstraction allows machine learning platforms to treat model artifacts as blobs created in the cloud and then shipped to devices, where they are then interpreted by the matching engine. In this way, the machine learning platform and its supported applications may be resilient to changes in the machine learning engine and/or may be agnostic or flexible to the particular engine or engine type used.

According to another aspect, a toolkit complementary to the on-device platform may provide a set of tools (eg, Python tools) for creating and simulating models in the cloud before they are delivered as artifacts to the device. In some implementations, the toolkit can generate from the same source artifacts (e.g., Python source artifacts) for different versions of machine learning engines or even different engine types (e.g., mobile-centric TensorFlow Lite vs. neural network libraries, etc.). .

In some implementations, the on-device machine learning platform may be implemented as an application or embedded in an application, such as a mobile application, for example. For example, in the context of the Android operating system, an on-device machine learning platform may be included in an Android Package Kit (APK) that may be downloaded and/or updated. In one particular example, the on-device machine learning platform may be included or implemented as part of a larger application that provides a number of different support services to other applications or to the device itself. For example, in addition to on-device machine learning platforms, larger applications enable computing devices to interact with digital distribution services (eg, downloading applications and/or updates from an “app store”) and/or other services. service can be provided. In another example, the on-device machine learning platform may be included or implemented as part of the device's operating system rather than as a standalone application.

The system and method of the present disclosure provide a number of technical effects and advantages. As one technical effect and advantage, the on-device machine learning platform enables the personalization of machine learning models based on locally stored device-specific training examples, leading to higher accuracy inference. Similarly, as described elsewhere herein, the on-device platform enables the device's participation in "federated learning" where local updates are aggregated to generate global updates, improving global model accuracy for all individuals. make it

As another exemplary technical effect and advantage, the on-device machine learning platform may safely include context signals in training examples and/or inference inputs. That is, contextual features can be added to training examples or inference inputs in a way that preserves privacy and compiles user-defined permissions. The inclusion of contextual information may improve the accuracy of the inference provided by the machine learning model.

As another example of technical effectiveness and benefits, on-device machine learning platforms can provide centralized services so that applications do not need to manage (e.g., train and/or run) machine learning models or interact with machine learning engines. can Thus, a given application does not need to store, manage, train and/or implement the machine learning model(s), but may instead communicate with the on-device machine learning platform to request and receive inferences from the model(s). This can reduce the data size of the application. It also simplifies the development and deployment of applications or other clients as application developers do not have to learn the complexities of different machine learning engines, but can instead simply rely on using the platform APIs.

Similar to the previous effects and benefits, an on-device machine learning platform may also enable easy updating of a single centralized service rather than all applications. For example, when a new version or type of machine learning engine is launched, typically only the on-device platform is updated to interact with the new engine, as applications or other clients use it on behalf of the platform rather than interacting with that engine. Should be. This eliminates the need to constantly ensure that your application is compatible with the latest version(s) of the machine learning engine, as you can rely on the on-device platform to keep you up to date as engine technology evolves.

As another exemplary technical effect and advantage, the on-device machine learning platform can improve communication network efficiency and usage. That is, under the past paradigm where machine learning is performed by servers rather than on-devices, various types of information (e.g., input data, training examples, inferences, model parameters, etc.) ) to the device by the server. However, as the present disclosure enables on-device prediction, training, example collection, and/or other machine learning tasks or functions, such information need not be transmitted (at least in all instances) over a communication network. Thus, communication network traffic, efficiency and usage are improved. In addition, since input data, training examples, etc. are not transmitted and received with the server, data security can be improved.

Referring now to the drawings, exemplary embodiments of the present disclosure will be discussed in more detail.

Exemplary devices and systems

1 shows a block diagram of an example computing device 102 that includes an on-device machine learning platform 122 in accordance with an exemplary embodiment of the present disclosure.

Computing device 102 may be any type of computing device including, for example, a desktop, laptop, tablet computing device, smart phone, wearable computing device, game console, embedded computing device, or other form of computing device. Accordingly, in some implementations, computing device 102 may be a mobile computing device and/or a user computing device.

Computing device 102 includes one or more processors 112 and memory 114 . The one or more processors 112 may be any suitable processing device (e.g., processor core, microprocessor, ASIC, FPGA, controller, microcontroller, etc.), and may be one processor or multiple processors operatively coupled. have. Memory 114 may include one or more non-transitory computer-readable storage media such as RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, and combinations thereof. Memory 114 may store data and instructions that are executed by processor 112 to cause computing device 102 to perform operations. Computing device 102 may also include a network interface 116 that enables communication over one or more networks (eg, the Internet).

The on-device machine learning platform 122 may enable on-device prediction, training, example collection, and/or other machine learning tasks or functions, which may be collectively referred to as “machine learning functions.”

The on-device machine learning platform 122 may be in the form of one or more computer programs stored locally on the computing device 102 (eg, a smart phone or tablet), which, when executed by the device 102 , and enable execution of on-device machine learning functions on behalf of the above locally stored applications 102a - c or other local clients. At least some of the on-device machine learning functions may be performed using one or more machine learning engines 128 implemented locally on the computing device 102 . The performance of on-device machine learning functions on behalf of one or more locally stored applications 120a - c or routines (which may be referred to as “clients”) is performed via one or more application programming interfaces (APIs) for on-device machine learning. It may be provided as a centralized service to clients capable of interacting with the platform 122 .

Further, in some implementations, the on-device machine learning platform 122 may include a context provider that securely inserts context features into the collected training examples and/or client-supplied input data used to generate predictions/inferences. Thus, on-device machine learning platform 122 may enable centralized training example collection, model training, and use of machine learning models 132a-c as a service to applications 120a-c or other clients. .

More specifically, computing device 102 may store or include one or more applications 120a - c (eg, mobile applications). Computing device 102 may also include and implement an on-device machine learning platform 122 and one or more machine learning models 132a - c . For example, machine learning models 132a - c may be stored by device 102 in a centralized model repository 130 managed by platform 122 .

In accordance with one aspect of the present disclosure, applications 120a - c communicate with on-device machine learning platform 122 via an API (which may be referred to as a “prediction API”) to provide input data, and one or more machines Predictions may be obtained based on input data from the learning models 132a - c. As an example, in some implementations, given a prediction plan (eg, instructions to run a model to obtain inference/prediction) and URIs for model parameters, the on-device machine learning platform 122 can download URI contents (e.g., prediction plans and parameters) and obtain one or more inferences/predictions by executing can Additionally, platform 122 may cache content (eg, in storage 130 ) for use in subsequent prediction requests.

Accordingly, the on-device machine learning models 132a - c may be accessed by the applications 120a - c by communicating with the on-device machine learning platform 122 through a client/service relationship. For example, individual machine learning models 132a - c may be provided to each application 120a - c and managed by the platform 122 . In another implementation, two or more applications 120a-c may share a single machine learning model 132a-c or a single application 120a-c may have two or more models 132a-c.

In some implementations, machine learning platform 122 may be a standalone multi-tenant service that may be referenced by applications 120a - c. As such, a given application 120a-c is not required to store, manage, train, and/or implement machine learning model(s) 132a-c, but instead simply interact with the on-device machine learning platform 122 . Communicate to request and receive inferences from model(s) 132a-c.

According to another aspect of the present disclosure, computing device 102 may further include a centralized example database 124 that stores training examples received from applications 120a - c. In particular, the on-device machine learning platform 122 may receive training examples from the applications 120a - c via an API (which may be referred to as an "aggregation API"), and may receive training examples from the centralized examples database 124 . You can manage the storage of examples. For example, each application 120a - c that is a client or tenant of the platform 122 may have its own collection(s) of examples stored in a centralized example database 124 , and the collection(s) are online may be supplemented and/or administered in this manner.

In some implementations, the on-device machine learning platform 122 evaluates each training example received from the application 120a-c according to one or more optional parameters associated with the application 120a-c providing the training example (e.g., For example, in its corresponding collection). As an example, the optional parameters may include a validity period parameter that defines a period of time during which training examples are stored (eg, deleted thereafter). In some implementations, optional parameters may be predefined and/or adjusted via commands provided to the platform 122 via the aggregation API.

According to another aspect of the present disclosure, in some implementations, the on-device machine learning platform 122 can safely insert contextual features that describe the context associated with the computing device 102 into training examples. For example, upon receiving training examples from applications 120a - c, the context provider component of on-device platform 122 determines one or more context characteristics and combines these context characteristics with training examples into a centralized example database 124 . ) can be stored in For example, contextual functions and data provided in a new training example may be stored as a single database item. Specific contextual characteristics that are then inserted or otherwise associated and/or stored with a training example received from a particular application 120a-c may be specified by optional parameters for that particular application 120a-c. As described above, these optional features can be customized or predefined via the ingest API. Accordingly, applications 120a - c can control which context features or context types are inserted into its training examples (eg, by defining optional parameters).

In some implementations, context features are inserted on the service side so that context features do not have to be available directly to applications 120a - c. In particular, in some implementations, the centralized example database 124 is not directly accessible by one or more applications 120a-c, so that contextual information stored with a particular training example is not accessible even by the application 120a-c that provided the training example. can't access

In some implementations, context characteristics may be grouped or classified according to a number of different context types. In general, each context type may specify or contain a well-known name and a set of context characteristics of the well-known type. One example context type is device information including example context characteristics such as audio status, network status, power connection, and the like.

In some implementations, the context provider requests the inserted value for a given context characteristic from the device (eg, the device's context manager) at insertion time/point. Alternatively or additionally, a context provider may register as a listener for one or more context updates, and may maintain a context feature cache of current values for context features based on one or more context updates. Then, when the context feature is inserted, the context provider accesses the context feature cache and inserts the current value held in the cache for that particular context feature.

In addition to or alternatively to inserting the context feature into the training example at storage time, the context provider may also perform the inserting of the context feature at inference time. In particular, when a particular application 120a-c or another client requests that inferences be generated based on some client-supplied input data (eg, via a prediction API), similar to the process described above for training example collection, , the context provider may insert or provide supplemental context features for input into the corresponding machine learning models 132a-c along with the input data. Accordingly, inferences may be made based at least in part on context information in addition to client-provided input data, which may help to improve the accuracy of the inferences.

According to another aspect, certain applications 120a - c or other clients (eg, as defined or controlled by the device user) may have access to only certain context features or context types. Therefore, the context provider can perform client privilege control. In particular, in some implementations, the on-device machine learning platform 122 or other device component may maintain a mapping of which clients have permission to access which context types or context features. When context features are inserted (eg, as a training example for storage or to supplement client-provided-input data at inference time), the context provider responds to the corresponding application 120a- for the context feature or context type to be inserted. c) Alternatively, you can check the authorization status of other clients. For example, a permission state for a particular application 120a-c and context type may describe whether such application 120a-c has permission to access that context type. The context provider inserts only those context features contained in the context type that the application 120a-c has permission to access, thereby providing context features/that the application 120a-c does not have permission to access (even in an indirect way). It will prevent you from accessing the type.

Similar to the validity period option parameter described above, in some implementations, each context feature may have an expiration period associated with or assigned to it. This expiry period information may be associated with each training example with context feature(s) included. In some implementations, at the end of an expiration period for a particular context feature provided in a particular training example, the value for that context feature may be deleted or removed from that training example. Alternatively, the entire training example may be deleted or removed.

Further, in some implementations, in response to a change to the authorization status for a particular application 120a-c or other client for a particular context characteristic or context type, the on-device platform 122 may Any value or item for such context feature(s) or type(s) associated with the training example associated with may be deleted. Also, in some implementations, the corresponding model(s) 132a - c may be retrained on the remaining data after deletion of the context feature value(s).

According to another aspect of the present disclosure, applications 120a - c communicate with an on-device machine learning platform via an API (also referred to as a “training API”) to learn machine learning based on training examples stored in a centralized database of examples. It can cause retraining or updating of the model. As an example, in some implementations, given a URI to a training plan (eg, an instruction for training a model), the on-device machine learning platform 122 may be configured based on previously collected examples (eg, may execute training on models 132a-c (by interacting with machine learning engine 128 to cause training of models 132a-c by engine 128). For example, training may be performed in the background at scheduled times and/or when the device is idle.

After retraining the models 123a-c, the retrained models 123a-c may be used to provide inferences as described elsewhere in this disclosure. In general, such inferences will have higher accuracy because models 123a-c have been retrained on user-specific data. Accordingly, on-device machine learning platform 122 may enable centralized example data collection and corresponding personalization of machine learning models 132a-c as a service to applications 120a-c or other clients.

According to another aspect, in some implementations, machine learning platform 122 may upload logs or other updates regarding machine learning models 132a - c to the cloud for detailed analysis of machine learning metrics. As an example, in some implementations, the on-device platform 122 configures parameters of the retrained machine learning models 132a-c or machine learning models 132a-c generated during the retraining of the models 132a-c. An update that describes a change to a parameter (eg, a “gradient”) may be determined. Platform 122 may send updates to a central server computing device (eg, the “cloud”) for integration with other updates provided by other computing devices. Thus, the platform may engage in a process known as “federated learning,” in which the device determines local updates to models 132a-c based on locally stored data, and then aggregates those local updates for aggregation (e.g., For example, it generates global updates to its models 132a-c by forwarding it to a cloud service (in a privacy-preserving and communication-efficient manner).

According to another aspect, in some implementations, in order to protect applications 120a - c from each other, each application 120a - c has specific functions provided by the on-device platform 122 (eg, You can have your own enclave for all functions). For example, the platform 122 may authenticate the applications 120a - c before the interface for accessing the platform 122 is returned to the applications 120a - c through the factory. The returned interface may represent the sole view of the application enclave on the platform 122 . In one example implementation of this process, when an application 120a-c connects to the API of the platform 122 , the application 120a-c receives a signed package token that verifies the identity of the application 120a-c. can provide Applications 120a-c cannot obtain the API interface without passing this authentication.

In some implementations, the enclave of each application within the platform 122 is account independent. Accordingly, multiple accounts associated with the same user profile on computing device 102 may share the same training data and status. This reflects that, in most cases, multiple accounts are for the same user, and different users of the computing device use different user profiles instead.

In some implementations, for certain functions (eg, access to context), permissions are required. As such, in some implementations, applications 120a - c wishing to use a particular context on the platform 122 have the right to access the particular context, even if they do not directly contact the context because the context is maintained within the platform 122 . has In some implementations, all relevant privileges may be passed to the platform call after being verified at the client, causing the platform 122 to logically operate with this privilege set(s). In some implementations, the platform 122 may request the user consent to the platform 122 with access to all permissions. In some cases, the context may also require a specific user to log in. Such a user can be specified by the application for that case, or it can be specified by the select (enemy) field in the options for context insertion. However, in some implementations, the user may not be automatically detected by the platform 122 . In some implementations, the API itself does not require authentication with that particular user account.

According to another aspect of the present disclosure, in some implementations, the on-device machine learning platform 122 may be completely abstracted from the underlying machine learning engine 128 . For example, machine learning engine 128 may be a tensorflow engine, neural network library, or other engine that enables implementation of machine learning models 132a - c for inference and/or training. This abstraction allows the machine learning platform 122 to treat model artifacts 132a-c as blobs that are created in the cloud and then delivered to a device (e.g., via dynamic model downloads), Here it is then interpreted by the matching engine 128 . In this way, the machine learning platform 122 and its supported applications 120a - c may be resilient to changes to the machine learning engine 128 and/or dependent on a particular engine 128 or engine type used. can be independent or flexible.

According to another aspect, a toolkit complementary to the on-device platform 122 may provide a set of tools (eg, Python tools) for creating and simulating models in the cloud before being delivered as artifacts to devices. In some implementations, the toolkit can generate from the same source artifacts (e.g., Python source artifacts) for different versions of machine learning engines or even different engine types (e.g., mobile-centric TensorFlow Lite vs. neural network libraries, etc.). .

In some implementations, the on-device machine learning platform 122 may be incorporated into or implemented as an application, such as, for example, a mobile application. For example, in the context of the Android operating system, the on-device machine learning platform 122 may be included in an Android Package Kit (APK) that may be downloaded and/or updated. In one particular example, the on-device machine learning platform 122 may be included or implemented as part of another application 120a - c or a larger application that provides a number of different support services to the device 102 itself. . For example, in addition to the on-device machine learning platform 122 , the larger applications may be that the computing device 102 provides a digital distribution service (eg, downloading applications and/or updates from an “app store”) and/or other services. You can provide services that allow you to interact with In another example, the on-device machine learning platform 122 may be included or implemented as part of the operating system of the device 102 rather than as a standalone application.

2 depicts a graphical diagram of an exemplary machine learning model deployment in accordance with an exemplary embodiment of the present disclosure. In particular, the application developer 202 may interact with the toolkit to create and test the model 204 . The model may be partitioned or represented at least in part by an inference plan 206 and a training plan 208 .

A “plan” may include a graph (eg, a tensorflow graph) and a protocol buffer (AKA “protobuf”) containing instructions on how to execute the graph. As an example, a plan may be a declarative description of a sequence of operations to be performed on a graph (eg, a tensorflow graph) that embeds (contains) the graph itself. A plan may describe how to query a collection for training data, how to feed it to a graph, and/or how to generate and deliver output.

Figure 2 shows two alternative (but optionally complementary) arrangements of arrangements. In a first deployment scheme, both the inference plan 206 and the training plan 208 are deployed on the cloud server 210 . The cloud server 210 provides the inference plan 206 and the training plan 208 to the device 214 .

Device 214 may implement inference plan 206 to generate inference. Device 214 may alternatively or additionally implement training plan 208 to perform on-device training based on locally stored data, also referred to as “personalization” or “personalized learning”. can

In a second deployment scheme, the inference plan 206 is deployed on the cloud server 210 as described above. The cloud server provides the inference plan 206 to the device 216 . Device 216 may implement inference plan 206 to generate inference.

However, in addition to or alternatively to deploying the training plan 208 on the cloud server 210 , in a second manner the training plan 208 is deployed on the federated server 212 . The federation server 212 provides the training plan 208 to the device 216 . Device 216 may implement training plan 208 to perform on-device training based on locally stored data. After this on-device learning, the device 216 may provide an update to the federation server 212 . For example, an update may describe one or more parameters of the retrained model or one or more changes to parameters of the model that occurred during retraining of that model.

The federation server 212 may receive many of these updates from multiple devices, and may aggregate the updates to generate an updated global model. The updated global model may then be sent back to device 216 .

Further, in some implementations, device 216 (eg, along with a toolkit) logs 218 or other updates regarding the machine learning model that can be used by developer 202 to obtain detailed analysis of machine learning metrics. can provide more. In some implementations, example metrics that may be calculated based on logs 218 include plots of check-in request results, traffic (eg, volume), loss and accuracy model metrics, phase duration, or other metrics. ), graphs or visualizations.

3 depicts a graphical diagram of an exemplary personalized and federated learning data flow in accordance with an exemplary embodiment of the present disclosure.

More specifically, FIG. 3 shows three different training data flows that may in some cases be used in a complementary manner. In a first data flow shown mainly in dashed lines at the bottom of FIG. 3 , training data is generated at the user device. After the training data is uploaded to a central authority, the machine learning model is trained or retrained based on the uploaded data. The model is then sent to the user device for use (eg, on-device inference).

In a second data flow, which may be referred to as personalization or personalized learning, training data generated on the user device is used to train or retrain the model on the device. The retrained model is then used by such devices. This personalized learning allows device-specific models to be trained and evaluated without centralized data collection, improving data security and user privacy.

In a third data flow, which may be referred to as federated learning, training data generated on the user device is used to train or retrain the model on the device. Therefore, since the actual user-specific training data is not uploaded to the cloud, data security and user privacy are improved.

After this on-device learning, the user device can provide an update to the central authority. For example, an update may describe one or more parameters of the retrained model or one or more changes to parameters of the model that occurred during retraining of that model.

A central authority may receive many of these updates from multiple devices, and aggregate the updates to create an updated global model. The updated global model can then be sent back to the user device. This approach may allow cross-device models to be trained and evaluated without centralized data collection.

4 depicts a block diagram of an exemplary on-device machine learning platform in accordance with an exemplary embodiment of the present disclosure. The platform illustrated and described with reference to FIG. 4 is provided as one example implementation only. Many other implementations of the on-device machine learning platform described herein are possible. The exemplary on-device machine learning platform may include or implement a main process 402 and a background process 404 .

The main process 402 can handle all API requests. The main process 402 includes a collection API service 420 that provides a training example collection service through the collection API 410; It includes a prediction API service 422 that provides an inference generation service through the prediction API 412 and a training API service 424 that provides a model training service through the training API 414 .

In some implementations, the aggregation API service 420 may maintain training examples with an automatic retention policy. In some implementations, the training API service 424 may draw data from the example collection and automatically execute training sessions at scheduled times and conditions as an invisible process. In some implementations, the prediction API service 422 may enable clients to execute inferences based on a given model, which may originate from a trainer or external sources.

In some implementations, the background process 404 may only host training and other periodic maintenance tasks. It is transactional and can be designed to be decomposed. In some implementations, the background process 404 obtains its state solely from the main process 402 .

As discussed further below, the context provider 430 inserts context information into the examples for both the ingest API service 420 and the prediction API service 422 . Storage component 440 may enable and perform storage of examples as well as bookkeeping state (eg, in centralized examples database 124 ). It can be based on LevelDB.

Depending on the prediction plan type, multiple prediction engines 432 may be accessed by the prediction API service 422 . Model parameters as well as prediction and training plan are provided or managed by artifact manager 434 . Artifact manager 434 may support retrieving artifacts from cloud server 210 , application assets, and/or files. It can also support mutable artifacts, for example to store training results used by predictors or other trainers.

The background process 404 may host a number of training engines 406 selected based on the training plan. For federated learning, the background process 404 may communicate with the federated learning server to upload accumulated training results using privacy protection techniques (eg, secure aggregation).

The log manager 444 may upload a log related to the machine learning model to the cloud server 210 for detailed analysis of the machine learning metric.

More specifically, in some implementations, the ingestion API service 420 stores training examples in a centralized example database 124 for later retrieval by the background process 404 (eg, to perform background training). may be a facility that permits, manages and/or performs For example, aggregation API service 420 may interact with storage component 440 to manage storage of training examples in centralized examples database 124 .

The collection API 410 may be simple. Once the client is authenticated, it can access the object as described in the example code below (where Task is an access method representing an asynchronous API call, and Task <Void> can be ignored or listened for error observation).

Learning.getCollectionClient(options) can allow access and configuration of exampleCollection. The 'option' parameter may include at least a collection name. If no additional options are provided, the default or previously configured options may be used. Exemplary options include the validity period of the content and context that must be inserted into the learning event before being stored in the database 124 .

CollectionClient.add(example) can add a new example to the repository.

CollectionClient.clear() may allow resetting of the contents of a collection.

Training API 414 and corresponding training API service 424 may schedule background process 404 to perform training. The background process 404 may implement or interact with one or more training engines 406 to pull data from the example collection and execute a training plan. A plan may be a declarative description of a sequence of operations to perform on a graph (eg, a tensorflow graph) that also embeds the graph itself. A plan may describe how to query a collection for training data, how to feed it to a graph, and/or how to generate and deliver output.

Each training plan type may be associated with a training plan engine 406 . Thus, an on-device machine learning platform can be extended by new types of schemes, thus representing any kind of machine learning execution that fits the general model of background training.

As an example, an example API for a trainer is provided below.

Learning.getTrainerClient(options) may take an option containing at least the trainer session name, and may create or reconfigure the training session. The session name can be an application-chosen constant, similar to the package name, and the session itself can be persistent. Options may specify the plan type, how the plan is obtained, and parameters specific to the plan type. Plans can be obtained in different ways depending on the plan type. For example, in the case of federation, the plan may be downloaded from the federated learning server, and in the case of personalization, it may be included in the asset or downloaded from the cloud server 210 .

TrainerClient.start(schedule) may start a training session at interface creation time and based on options passed to a given schedule. The schedule may be continuous or one-time. In some implementations, in both cases, training is scheduled only if device conditions permit. For example, in some implementations, training will be scheduled or performed only when the device is idle or charging.

TrainerClient.stop() may allow cancellation and/or removal of training sessions.

Prediction API 412 allows clients to supply inputs based on a trained model and derive predictions therefrom. Like trainers, predictors can, in some implementations, be plan-driven. The plan is a declarative description of the operations to be performed on the graph and how to obtain an input and produce an output.

As an example, exemplary prediction API code is as follows.

Learning.getPredictorClient() can return a predictor based on the given options. Options may specify how to obtain plan and model parameters for prediction. Options may also specify context characteristics that should be automatically inserted into candidate examples before being passed to the prediction engine.

predictRank() may return a prediction for a ranking problem derived from a given context example and a specified candidate. Over time, additional application-specific prediction methods may be introduced.

The code below shows an example of use of the three exemplary APIs 410 , 412 and 414 described with reference to FIG. 4 .

First, configuration options may be defined. Typically these options may be obtained from the expression construct component 442 by the application, but may be defined as static constants for simplicity.

Note how URIs can be used to refer to artifacts that describe model parameters as well as training plans and prediction plans. Plans may encode information about a graph (eg, a tensorflow graph) and how to execute the graph. In some implementations, the plans may be generated by a tool included in a corresponding toolbox (eg, a Python tool). In some implementations, the model parameters may be an opaque representation of weights associated with the plan. The URI may refer to "model repository" (mrepo:), which means it is downloaded to the device (eg, from cloud server 210), but may also refer to a locally cached file (file:). For example, the artifact manager 134 may manage the download of model artifacts from the server 210 and/or other model management tasks.

For file artifacts, dependencies can be defined between APIs. For example, TrainerOptions may be defined to generate file artifacts with trained parameters used by the prediction API service. On-device machine learning platforms can handle input-output dependencies internally by deferring actions or rejecting them with an appropriate error code if the required input has not yet been generated.

By providing the above configuration, the platform can include some API code where training examples are continuously fed as a collection. For example, the following code could include an add training example.

Whenever an example is added to the cache, the context specified by COLLECTION_OPTIONS may be added as a feature. Users do not need to limit the size or lifetime of additional data they can process based on the options provided.

To schedule training, an application can ensure that background training is configured and scheduled, typically using the current options at the time of creation. In some implementations, if training is already scheduled and the configuration is not changed, the next example call is not affected.

Finally, other parts of the example code can utilize the training results using the prediction API. As an example, this can be seen as provided in the example code below.

As already mentioned, in some implementations, the context provider 430 of the on-device machine learning platform may insert the context feature into the learning event. This can happen on the service side, so there is no need to use the context directly in the application.

Two exemplary points in time at which context can be inserted are:

1. Before the example is saved in the example collection. The inserted context can be specified by CollectionOptions.

2. Before the example is passed to the prediction engine. The inserted context can be specified by PredictorOptions.

In general, each context category may specify a set of features with well-known names and well-known types appended to their instances (eg, TensorFlow example protos). The inserted value for a given context characteristic may be either requested from the system at the time of insertion, or it may be a cached value that is periodically updated by the platform's internal context provider 430 .

Examples of context characteristics include audio state; day attribute; calender; detection activity; user-specified place (eg, "home" versus "work"); network status; power connection; screen features; user location; user location prediction; WiFi scan information; weather, or other contextual characteristics.

Importantly, the training examples and context features described herein are provided simply to illustrate example data that may be stored with the training examples or used to provide inference by an on-device platform. However, if the user does not consent after obtaining information about what data the user collects and how such data is used, such data will not be collected, used or analyzed. In addition, a tool may be provided to the user to revoke or modify the privilege set. In addition, certain information or data may be processed in a number of ways prior to being stored or used so that personally identifiable information is removed or stored in an encrypted manner.

Since the on-device platform does not require any context itself, the context provider 430 collects the context for the client application. For example, a client application may need a "location" as a function for its machine learning model. Such an application can explicitly inform the on-device location platform that it needs a "location" context.

The on-device platform may first check whether the client application has permission to access the device location. Otherwise, the platform does not provide a context to the client. On the other hand, if the application has permission, the location context for that client will be populated when it sends training to the platform. Thus, in some implementations, in each instance the client sends a yes to the on-device platform, the platform checks their permissions to determine whether the client has permission to access the context in which they assert.

Note that, in some implementations, no actual contextual content is provided to the application. Instead, the context is simply populated with training examples for the client. The training examples are maintained within the on-device platform database, so the client has no access to the actual context content.

In some implementations, certain types of contexts require user accounts to access similar place aliases and calendars. In some implementations, the context provider 430 itself does not indicate which account to use for the context. In this case, the client application must specify an account. In some implementations, if no account is specified by the client, only the context that does not require an account will be provided to the client.

The context used by the on-device platform is typically provided by a context manager (eg, 126 as shown in FIG. 1 ). The context manager may be located in multiple places, including, for example, the platform, the applications comprising the platform, and/or the operating system of the device. In some implementations, to improve performance, the context provider 430 may register a listener with the context manager to always keep the latest context updates in on-device platform memory.

In some implementations, the on-device platform may perform or include stale context expiration. In some implementations, when users turn off a context signal on the device (eg, turn off location or activity recognition), the context manager does not inform the on-device platform that the context has been turned off. Instead, the context manager simply stops sending updates for these context(s) to the on-device platform. Thus, in order not to use the stale context for future events, the platform may cause the stale context to expire. In particular, based on context attributes, different expiry time periods may be defined for each context characteristic or context type. When a context reaches an expiration time period, it can be deleted.

Another example context feature includes a place alias context feature. In particular, whether the user is at home or at work is an important feature in many client applications. Considering that the user's home/work location does not change frequently, the platform may request the current home/work alias once the context provider 430 is configured. If the user has consented to the use of such information, the user's home/work location may be cached and the context provider 430 may use the location context to determine if the user is at home by comparing that location context to the cached location. Or you can decide whether you are at work. In some implementations, the venue alias information may be received from a context manager or a venue API.

A context manager can pass a context in at least two ways: In the first example, the on-device platform can register as a listener. In this case, the on-device platform may maintain a cache of updated contexts, which in some implementations means that the on-device platform is an always-on service. Listeners allow you to cache updated data.

Registering as a listener has the following advantages:

1. Low latency. All contexts are cached within the on-device platform and translated into a format suitable for machine learning.

2. If the Inter Process Call (IPC) to the on-device platform is high-speed, the caching context saves battery.

In the second example, the on-device platform can obtain the current context in one shot. In this case, the context manager can provide another API to get all the current contexts in one shot. When this API is used, the on-device platform generally does not maintain a cache of contexts, but fetches the current context on request.

In this mode, the context manager is asked to maintain an updated context for the platform, so the platform acquires additional UDC privileges that are not normally required for the first option.

Advantages of one-shot mode include saving battery when IPC to the on-device platform is slow.

In some implementations, the on-device platform may acquire user rights for all contexts listed above, but certain clients may not have the same rights as the on-device platform. Thus, the platform can control the permissions on the client. In some implementations, this can be done using a package manager to extract the permissions associated with the client application. The platform may maintain a mapping between contexts to rights. Thus, clients typically explicitly request the context they want to use when registering themselves with the on-device platform. The platform checks whether the client has permission to access the requested context. Only the context in which the application has privileges is used to train and infer that model.

In addition, the rights of the client can be changed immediately. For example, a user can authorize an application to use their location when the application finds it useful, and then revoke the permission. To handle this case, whenever a new event comes from a client, the platform may check their existing privileges and include the event to be trained on and the corresponding context.

During inference, the platform may also analyze current client privileges to allow use of contextual features for prediction. As such, in some implementations, machine-learned models may accommodate missing features for training and inference.

According to another aspect, in some implementations, to use the on-device platform API, clients may add an API key to their application. For example, an application can be signed with a digital certificate in which the client holds a private key. An application can obtain a key by registering with the central authority that manages the API.

One example authentication procedure may include:

1. The on-device platform obtains the package name and API key pair when the client registers with the on-device platform. The on-device platform sends the package name and API key pair to a central authority for verification.

2. Once the client is verified, the on-device platform will generate a platform key for the client to use on the device.

3. For future API calls on the device, the client must provide the platform key to the on-device platform.

[159]

The on-device platform may check whether the platform key matches the package name and API key. However, the on-device platform generally does not send it to a central authority for re-verification.

In some cases, a device may be shared by more than one user. As such, the on-device platform can extract the primary account from the list when the API is used. The on-device platform can link training data to the primary account and update the relevant model. If the account's model is not available on the device, the on-device platform can either download the client application's model from the cloud or use a base model (eg, the average model in federated learning).

According to another aspect, in some implementations, a user may clear his location history or account history. In such a case, the on-device platform may remove all corresponding contexts. Also, in some implementations, the platform may retrain the model using the remaining context for this user in this case.

By way of example, FIGS. 5A and 5B show block diagrams of an example machine learning platform 122 that embeds context features in accordance with example embodiments of the present disclosure. In particular, Figure 5a shows a context provider 430 inserting context features before training examples where examples are stored in an example collection, and Figure 5b inserting context features before examples are passed to a prediction engine implementing a machine learning model. Context provider 430 is shown.

6A and 6B show block diagrams of an exemplary device for performing model training in accordance with an exemplary embodiment of the present disclosure. In particular, Figures 6a and 6b show the background training process.

As shown in FIG. 6A , the application process may supply training data to a training data store. The training process may be scheduled in the background (eg, when allowed by certain device conditions such as idle and/or plug-ins).

The training process selects the model state and training plan, iteratively selects data from the cache to train the model, and finally publishes the statistic(s) and model update message(s) (e.g. to the cloud server). can do. In some implementations, since the training phase can be lengthy (eg, several minutes), the training process can be stopped and resumed based on changes in device state.

Figure 6b is similar to Figure 6a except that it includes a federated learning service that enables federated learning. As shown in FIG. 6B , the training plan and model may be distributed by a federated learning service. The training process may perform training in the background to generate model updates. Model updates can be uploaded to the federated learning service (for use in aggregation). Also, in some implementations, quality assurance (eg, semi-automated quality assurance) may extract a learned model and redeploy the model to a device.

7 shows a graphical diagram of an exemplary federated learning process in accordance with an exemplary embodiment of the present disclosure.

Exemplary method

8 depicts a flow diagram of an example method 800 for generating inferences using a machine learning model in accordance with an exemplary embodiment of the present disclosure.

At 802 , the computing system may receive input data from the first application via the predictive application programming interface.

At 804 , the computing system may supplement the input data with context features requested by the first application and to which the first application has access. In some implementations, at 804 , the computing system can determine a permission status for the first application for each of the one or more context types. In some implementations, the input data is supplemented with only context features included in the context types to which the first application has access.

At 806 , the computing system may generate, using the first machine learning model, at least one inference based at least in part on the input data and at least in part on the additional context characteristics. At 808 , the computing system may provide the at least one inference to the first application via the predictive application programming interface.

9 depicts a flow diagram of an example method 900 for collecting training examples for performing machine learning in accordance with an exemplary embodiment of the present disclosure.

At step 902 , the computing system may receive a new training example from the first application via the aggregation application programming interface. At step 904 , the computing system may add a new training example with context features requested by the first application and to which the first application has access. In some implementations, at 904 , the computing system can determine a permission status for the first application for each of the one or more context types. In some implementations, the new training example is added only with context features included in the context types to which the first application has access.

At step 906 , the computing system may store the new training examples along with the context features in a centralized example database. In some implementations, storing the new training examples in step 906 includes storing the new training examples in a centralized example database according to one or more optional parameters predefined for the first application through the ingest application programming interface. may include As an example, the one or more optional parameters may include at least a validity period parameter defining a period of time for which training examples are stored.

In some implementations, storing the new training example at step 906 may include assigning an expiration period to at least a first one of the one or more context features. Method 900 may further include deleting the first context feature or the entire new training example from the centralized example database at the end of an expiration period assigned to the first context feature.

In some implementations, method 900 includes receiving an indication of a change to a permission status for a first application for at least one context type; and in response to the change to the permission status, deleting from the centralized example database any context characteristics of the at least one context type associated with the training example associated with the first application. Further, in some implementations, the method 900 includes, after deleting the context features, retraining one or more machine learning models associated with the first application using training examples associated with the first application in a centralized example database. may include more.

10 depicts a flow diagram of an exemplary method 1000 for training a machine learning model in accordance with an exemplary embodiment of the present disclosure.

In step 1002, the computing system is configured to receive instructions from the first application via the training application programming interface to retrain the first machine learning model based at least in part on one or more training examples stored by the centralized example database. can

At 1004 , the computing system may retrain the first machine learning model based at least in part on one or more training examples stored by a centralized example database.

At step 1006, the computing system may determine an update describing one or more parameters of the retrained model or one or more changes to one or more parameters that occurred during retraining of the model.

At step 1008 , the computing system may send the update to a central server computing device for aggregation of the update with other updates provided by other computing devices.

additional start

Techniques discussed herein refer to servers, databases, software applications and other computer-based systems as well as actions taken and information transferred to/from such systems. The inherent flexibility of computer-based systems allows for a wide variety of possible configurations, combinations, and distinctions between tasks and functions between and among components. For example, the processes discussed herein may be implemented using a single device or component or multiple devices or components in combination. Databases and applications can be implemented on a single system or distributed across multiple systems. Distributed components can operate sequentially or in parallel.

Although the subject matter has been described in detail with respect to various specific exemplary embodiments of the disclosure, each example is provided for illustrative purposes rather than limitations of the disclosure. Those of ordinary skill in the art, upon understanding the foregoing, can readily make changes, modifications, and equivalents to such embodiments. Accordingly, this disclosure is not excluded from covering such alterations, modifications and/or additions to the subject matter as would be apparent to those skilled in the art. For example, features shown or described as part of one embodiment may be used in conjunction with another embodiment to create another embodiment. Accordingly, this disclosure is intended to cover such modifications, variations and equivalents. In particular, although FIGS. 8-10 each show steps performed in a specific order for purposes of illustration and discussion, the methods of the present disclosure are not limited to the particularly illustrated order or arrangement. Various steps of methods 800, 900, and 1000 may be omitted, rearranged, combined, and/or adapted in various ways without departing from the scope of the present disclosure.

Claims (20)

A computing device comprising:
one or more processors; and
one or more non-transitory computer readable media, the non-transitory computer readable media comprising:
one or more applications implemented by one or more processors;
one or more machine learning models; and
store instructions that, when executed by one or more processors, cause the computing device to implement an on-device machine learning platform that performs operations;
The actions are
receiving input data from a first one of the one or more applications via the predictive application programming interface;
using at least a first of the one or more machine learning models to generate at least one inference based at least in part on the input data; and
and providing at least one inference generated by the first machine learning model to the first application via the predictive application programming interface. According to claim 1,
the one or more non-transitory computer readable media comprising:
further store a centralized example database storing training examples received from one or more applications; and
The operations performed by the on-device machine learning platform are:
receiving new training examples from the first application via the aggregation application programming interface; and
and storing the new training examples in a centralized example database. 3. The method of claim 2,
storing the new training examples in the centralized example database comprises storing the new training examples in the centralized example database according to one or more optional parameters predefined for the first application through an aggregation application programming interface;
wherein the one or more optional parameters include at least one time-to-live parameter defining a time period for which training examples are stored. 3. The method of claim 2,
the one or more applications include a plurality of applications; and
and the centralized example database stores training examples received from two or more different applications of a plurality of applications. 3. The method of claim 2,
The operations performed by the on-device machine learning platform are:
receiving instructions from the first application via the training application programming interface to retrain the first machine learning model based at least in part on one or more training examples stored by the centralized example database; and
responsive to the instruction, causing the first machine learning model to be retrained based at least in part on one or more training examples stored by the centralized example database. 6. The method of claim 5,
The operations performed by the on-device machine learning platform are:
after causing the first machine learning model to be retrained, using the retrained first machine learning model to generate at least one additional inference; and
and providing at least one additional inference generated by the retrained first machine learning model to the first application via the predictive application programming interface. 6. The method of claim 5,
the one or more non-transitory computer-readable media further storing a machine learning engine; and
causing the first machine learning model to be retrained based at least in part on one or more training examples stored by the centralized example database includes causing a machine learning engine to retrain the first machine learning model according to a training plan; Computing device comprising a. According to claim 1,
the one or more applications include a plurality of applications; and
The instructions, when executed by the one or more processors, cause a computing device to implement an on-device machine learning platform, comprising:
instructions that, when executed by one or more processors, cause a computing device to implement an on-device multi-tenant machine learning platform that performs operations comprising:
maintaining a plurality of independent enclaves each for a plurality of applications;
receiving from the first application a signed package token that verifies the identity of the first application;
authenticating the signed package token; and
responsive to authentication of the signed package token, providing access to a first enclave of a plurality of stand-alone enclaves, the first enclave being associated with a first application. According to claim 1,
The instructions that, when executed by the one or more processors, cause a computing device to implement an on-device machine learning platform comprising:
mobile application; or
A computing device comprising a portion of an operating system of the computing device. According to claim 1,
The computing device comprises a mobile computing device,
wherein the one or more applications include one or more mobile applications. One or more non-transitory computer-readable storage media that collectively store instructions that, when executed by one or more processors, cause a computing device to implement an on-device machine learning platform that performs operations comprising:
receiving input data from a first one of the one or more applications stored on the computing device via the predictive application programming interface;
using at least a first of the one or more machine learning models stored on the computing device to generate at least one inference based at least in part on the input data; and
and providing at least one inference generated by the first machine learning model to a first application via a predictive application programming interface. 12. The method of claim 11,
the one or more non-transitory computer-readable media further stores a centralized example database storing training examples received from one or more applications; and
The operations performed by the on-device machine learning platform are:
receiving new training examples from the first application via the aggregation application programming interface; and
and storing the new training examples in a centralized example database. 13. The method of claim 12,
storing the new training examples in the centralized example database comprises storing the new training examples in the centralized example database according to one or more optional parameters predefined for the first application through an aggregation application programming interface;
and the one or more optional parameters include at least one time-to-live parameter defining a time period for which training examples are stored. 13. The method of claim 12,
the one or more applications include a plurality of applications; and
and the centralized example database stores training examples received from two or more different ones of a plurality of applications. 13. The method of claim 12,
The operations performed by the on-device machine learning platform are:
receiving instructions from the first application via the training application programming interface to retrain the first machine learning model based at least in part on one or more training examples stored by the centralized example database; and
responsive to the instruction, causing the first machine learning model to be retrained based at least in part on one or more training examples stored by the centralized example database. 16. The method of claim 15,
The operations performed by the on-device machine learning platform are:
after causing the first machine learning model to be retrained, using the retrained first machine learning model to generate at least one additional inference; and
and providing at least one additional inference generated by the retrained first machine learning model to the first application via the predictive application programming interface. 16. The method of claim 15,
the one or more non-transitory computer-readable media further storing a machine learning engine; and
causing the first machine learning model to be retrained based at least in part on one or more training examples stored by the centralized example database includes causing a machine learning engine to retrain the first machine learning model according to a training plan; A non-transitory computer-readable storage medium comprising a. 12. The method of claim 11,
the one or more applications include a plurality of applications; and
The instructions, when executed by the one or more processors, cause a computing device to implement an on-device machine learning platform, comprising:
instructions that, when executed by one or more processors, cause a computing device to implement an on-device, multi-tenant machine learning platform that performs operations comprising:
maintaining a plurality of stand-alone enclaves each for a plurality of applications;
receiving from the first application a signed package token that verifies the identity of the first application;
authenticating the signed package token; and
responsive to authentication of the signed package token, providing access to a first enclave of a plurality of stand-alone enclaves, the first enclave being associated with a first application computer readable storage medium. 12. The method of claim 11,
The instructions that, when executed by the one or more processors, cause a computing device to implement an on-device machine learning platform comprising:
mobile application; or
A non-transitory computer-readable storage medium comprising a portion of an operating system of a computing device. A computer implemented method comprising:
receiving, by the computing device via the aggregation application programming interface, a new training example from a first one of the plurality of applications stored on the computing device;
storing, by the computing device, the new training example in a centralized example database;
Receive, by the computing device, instructions from the first application via the training application programming interface to retrain, by the computing device, a first machine learning model stored by the computing device based at least in part on one or more training examples stored by the centralized example database. to do;
in response to the instruction, causing, by the computing device, to retrain the first machine learning model based at least in part on one or more training examples stored by the centralized example database;
receiving, by the computing device, input data from the first application through the predictive application programming interface;
using the first machine learning model to generate at least one inference based at least in part on the input data; and
and providing at least one inference generated by the first machine learning model to the first application via a predictive application programming interface.

Images